[Bugfix][MiniCPM-o] Restore the duplex tests removed in #7413

[chickeyton]

What broke

Since #7413 a duplex session never reports the server-side per-response TTFT/TTFP that #7242 added for the MiniCPM-o Omni-DuplexEval / OmniInteract benchmarks. response.created and the audio deltas carry no response_request_metrics, so vllm_omni.clients.duplex.EventCollector and benchmarks/duplex_session_metrics.py silently fall back to the client's receive clock ("source": "client_monotonic_receive"), while docs/cli/bench/serve.md still promises "Per-response TTFT and TTFP start when the server begins executing the native model-turn request". The same commit's TPOT weighting fix was lost too: a one-token segment is weighted as one interval again and skews the response mean.

Repro

# on main: 4 of the 5 new tests fail (the methods and the wire fields do not exist)
pytest -q tests/engine/duplex/test_engine_session.py tests/engine/duplex/test_session_runner.py \
  -k "response_timing or response_tpot or response_request_metrics"

The strengthened e2e gate fails on main the same way: every response's request_metrics.source is client_monotonic_receive instead of server_request_start_and_client_receive.

CUDA_VISIBLE_DEVICES=0 pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py \
  tests/e2e/online_serving/test_duplex_client_live.py -m "core_model and cuda" --run-level core_model

Root cause

#7413 replaced entrypoints/duplex/{protocol,runtime_bridge,session_runner}.py with the engine-resident DuplexEngineSession / ModelChannel, and the port did not carry over mark_model_turn_request_started, mark_response_first_outputs, the response_request_metrics attachment, or the max(tokens - 1, 0) TPOT weight. The client side of #7242 survived, so the feature degraded silently instead of failing.

Fix

• engine_session.py: per-turn request-start ledger on ResponseState, mark_model_turn_request_started / mark_response_first_outputs, binding at begin_response, pruning at complete_model_turn, resets on end / barge-in / close, and the interval-weighted TPOT mean.
• model_channel.py: anchors the clock when the append is submitted and attaches the metrics to response.created (lands under response.metadata.duplex_event) and to every speak / audio delta (metadata.vllm_omni), the two places the client already reads.
• Tests: the three [Frontend][Benchmark] Add duplex performance metrics for OmniInteract and Omni-DuplexEval benchmark #7242 session tests plus a lifecycle test, a runner test over the real MiniCPM-o plugin, the MiniCPM-o plugin policy tests re-homed from the deleted tests/engine/duplex/test_duplex_runtime.py, and the duplex e2e gate now asserts the server-side anchor. The 11 pre-existing mypy-3.10 errors in the two touched test files are fixed so pre-commit passes.
• Docs: three passages still described the session-per-request DuplexChatCompletionsAdapter that was removed before [Core][Frontend] Unified Full-duplex Framework #7413 merged; they now match the chat route that exists.

Tests restored

Thirteen tests across four files: ten restore tests #7413 deleted, three are new.

From tests/entrypoints/openai_api/test_duplex_handler.py (deleted by #7413), now in tests/engine/duplex/test_engine_session.py

Original	Restored as
test_response_timing_binds_latest_request_start_for_model_turn	same name, on DuplexEngineSession
test_response_tpot_fallback_ignores_single_token_segment	same name
test_response_tpot_keeps_token_weighted_value_when_itls_exist	same name

From the same file, now in tests/engine/duplex/test_session_runner.py

Original	Restored as
test_continuous_response_metrics_accumulate_only_owned_model_units (the response_request_metrics half)	test_response_request_metrics_are_measured_from_the_model_turn_request_start, driven through the runner harness with the real MiniCPM-o plugin

From tests/engine/duplex/test_duplex_runtime.py (deleted by #7413), now in the new tests/model_executor/models/minicpmo_4_5/duplex/test_plugin_policy.py

Original	Restored as
test_minicpmo_extension_owns_stage_sampling_overrides	test_plugin_owns_stage_sampling_overrides_without_mutating_defaults
test_minicpmo_output_decision_uses_raw_streaming_token_snapshot	test_listen_decision_uses_the_raw_streaming_token_snapshot
test_minicpmo_output_decision_ignores_output_level_token_history (x2 params)	test_listen_decision_ignores_output_level_token_history
test_minicpmo_output_decision_uses_completion_token_ids (x2 params)	test_listen_decision_uses_completion_token_ids
test_minicpmo_output_decision_uses_completion_stop_reason	test_listen_decision_uses_completion_stop_reason
test_duplex_scheduler_token_budget_estimates_pcm_slots	test_scheduler_token_budget_estimates_pcm_slots
test_duplex_scheduler_token_budget_ignores_client_budget_fields	test_scheduler_token_budget_ignores_client_budget_fields

New, no deleted counterpart

• test_response_timing_is_empty_without_a_request_start_and_is_scoped_to_one_response (test_engine_session.py): lifecycle of the timing state across end, turn completion, and barge-in.
• test_a_speak_segment_is_not_a_listen_decision (test_plugin_policy.py): negative case for the listen decision.
• tests/e2e/online_serving/test_minicpmo_4_5_duplex.py: _assert_request_metrics now requires source == "server_request_start_and_client_receive" and the native request-start measurement origin, so the CI e2e gate catches this regression.

Deliberately not restored

• test_duplex_scheduler_token_budget_one_frame_is_one_block, ..._matches_stage0_hd_slicing, ..._unreadable_base_frame_keeps_hd_fallback and test_duplex_hd_slice_count_matches_official_grid from test_duplex_runtime.py: [Bugfix][MiniCPM-o] Size the duplex HD slice reservation from the frame #7654 re-covers them in tests/model_executor/models/minicpmo_4_5/duplex/test_scheduler_token_budget.py.
• The remaining 68 test_minicpmo_* handler tests, which targeted the serving architecture [Core][Frontend] Unified Full-duplex Framework #7413 removed.

Validation (1x NVIDIA L20X, vllm 0.29.0, torch 2.13.0+cu129)

Suite	Result
pre-commit run --files <changed> (all hooks incl. mypy-3.10)	passed
Duplex CPU sweep: tests/engine/duplex tests/clients tests/model_executor/models/minicpmo_4_5 tests/benchmarks/{test_omniinteract.py,patch,metrics} tests/entrypoints/duplex tests/engine/test_duplex_orchestrator.py tests/engine/test_duplex_omni_engine.py tests/entrypoints/test_duplex_omni.py tests/entrypoints/openai_api/test_duplex_api_server.py tests/entrypoints/openai/test_duplex_session_attachment.py tests/worker/test_native_duplex_hooks.py tests/protocol tests/engine/test_duplex_import_boundary.py tests/engine/test_output_processor.py	1477 passed, 5 skipped in 7m14s
MiniCPM-o duplex e2e, -m "core_model and cuda" --run-level core_model	2 passed, 11 deselected in 15m00s
New and restored tests (test_engine_session.py, test_session_runner.py, test_plugin_policy.py)	89 passed

Not restored on purpose: the frame-size-aware vision slot budget from #7271 was also dropped by #7413 but has since been re-fixed by #7654; the duplex Seed-TTS perf baselines removed in #7413 are left out because the native per-unit path changed.

🤖 Generated with Claude Code

[chickeyton]

Pre-check report (precheck-pr, full mode, bug-fix checklist)

Run after fixing the blocking issue found on the first pass: pre-commit's mypy-3.10 hook reported 11 pre-existing type errors in the two touched test files (tests/engine/duplex/test_session_runner.py: object typed harness output / event; tests/e2e/online_serving/test_minicpmo_4_5_duplex.py: dict[str, object] result indexed as str / int / list). Fixed by typing the harness fakes as SimpleNamespace / DuplexEvent and giving the seeded-turn helper a TypedDict result.

Pre-check report for duplex_restore_tests

  Mode: full
  Type: bug-fix

  Dimension            Result
  ───────────────────  ──────
  PR title format      ✓ [Bugfix][MiniCPM-o] prefix, no WIP
  Code quality         ⚠ 11 new SimpleNamespace fakes in tests (duck-typed stage
                         outputs and a clock stub, same convention as the existing
                         runner harness); no new **kwargs lookup, broad except,
                         Any hint, hot-path copy, or event-loop blocking
  Examples policy      ✓ no new Python example paths
  Simplification       ✓ no candidates (two new session methods, both called)
  PR body 4 sections   ✓ what broke / repro / root cause / fix
  Fix is minimal       ⚠ three doc passages about the removed session-per-request
                         chat adapter are corrected alongside the fix
  Repro is runnable    ✓ pytest -k "response_timing or response_tpot or
                         response_request_metrics": 4 failed on main, 5 pass here
  Branch is rebased    ✓ merge-base == origin/main (573ec4cd)
  Regression test      ✓ session, runner (real MiniCPM-o plugin) and e2e assertion
  Root cause match     ✓ restores exactly the #7242 semantics the port dropped
  No silent failure    ✓ no new except / fallback
  Upstream pattern     ✓ same methods and wire placement the client already reads
  Environment          — not version-dependent
  Pre-commit gates     ✓ all hooks pass on Linux after the mypy fix (SPDX,
                         forbidden imports, torch.cuda, test marks, markdownlint,
                         ruff, typos, mypy-3.10)
  Dead code            ✓
  Test coverage        ✓ engine/ change covered by unit, runner and e2e tests
  Import hygiene       ✓ one new import (TypedDict), used
  Unrelated changes    ⚠ see "Fix is minimal"

  Verdict: 0 blocking | 2 warnings | ready for review

Evidence: pre-commit PRECOMMIT_EXIT=0; duplex CPU sweep 1477 passed / 5 skipped; MiniCPM-o duplex e2e (core_model, 1x L20X) 2 passed.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/engine_orchestration.md, docs/design/module/observability.md.

Module owners: @tzhouam @fake0fan @lishunyang12

Routing: @tzhouam via module of the changed files, CODEOWNERS; @fake0fan via module of the changed files; @lishunyang12 via module of the changed files

@chickeyton, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[vllm-omni-review-bot]

Omni ReviewBot triage note

Automated triage of commit e3f0acf4776d produced:

• Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

[amy-why-3459]

@yenuo26 PTAL

[vllm-omni-review-bot]

Omni ReviewBot: no human activity for 7 days

@chickeyton this pull request has had no human commit, comment or review since 2026-09-20. Please confirm the current plan and next step. The author or a maintainer decides whether to change the PR state.

To keep it moving, any one of these is enough: push an update, reply to the open blocker, or post the current plan and timeline.

[vllm-omni-review-bot]

Omni ReviewBot: no human activity for 14 days

@chickeyton this pull request has had no human commit, comment or review since 2026-09-20. Please consider marking this PR as draft until work can resume. The author or a maintainer decides whether to change the PR state.

To keep it moving, any one of these is enough: push an update, reply to the open blocker, or post the current plan and timeline.

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Strict on cursor (cursor-grok-4.6-high) under experiment fleet-strict-cursor-grok46-zcode-glm53flash-5050-c5-z10-20261002.